NVIDIA is expanding the core value proposition of its next-generation Vera Rubin platform from inference cost to model training efficiency, betting on the new metric of "intelligence per dollar" to position continuous post-training as the central compute demand in the era of Agentic AI.
In a corporate blog post, NVIDIA detailed that with the rise of Agentic AI, model post-training has evolved from a one-time final step into a continuous, cyclical core workload. Unlike traditional generative models, agentic models require planning, tool calling, and autonomous error correction during operation, and their environments can change weekly, leading to an ongoing accumulation of post-training compute needs. NVIDIA stated that the Vera Rubin platform is co-designed for this workload, requiring only a quarter of the GPUs of the previous-generation Blackwell platform when training the largest models.
Shifting Compute Demands for the Agentic Era
This statement directly relates to NVIDIA's compute sales logic: a never-ending post-training cycle means client demand for GPU clusters will shift from project-based to continuous, expanding the potential market size. Companies like Prime Intellect, Perplexity, and Together AI, which already run post-training workloads on NVIDIA's platforms, have indicated plans to migrate or expand to the Vera Rubin platform.
Post-Training as the Core Compute Driver
NVIDIA systematically outlined the strategic importance of post-training in its blog. The pre-training phase gives a model language fluency, while true "intelligence"—including writing code, planning multi-step tasks, using search tools, and recovering from errors—is formed during post-training.
Post-training employs reinforcement learning (RL) techniques: the model generates attempts for a given task (forward propagation), which are scored and then used to update the model's weights (backward propagation). Through millions of iterations, the model's capabilities gradually improve. NVIDIA noted that this process is extremely compute-intensive, requiring thousands of environments to generate rollouts in parallel while keeping accelerators fully utilized.
NVIDIA positions "intelligence per dollar" as a higher-level metric than "cost per token": the former measures the operational efficiency of an inference factory, while the latter assesses whether the investment to build and continuously maintain a deployable model is worthwhile. The two are interlinked—lowering the cost per token also reduces the cost of building model intelligence, and higher model intelligence increases the service value of each token.
Nemotron Ultra Provides a Verifiable Benchmark
To support these claims, NVIDIA disclosed post-training details for its open-weight model, Nemotron 3 Ultra. This 550-billion parameter model uses a Mixture of Experts (MoE) architecture, and its full post-training pipeline runs on the NeMo RL framework.
On the SWE-bench Verified benchmark, which tests real-world programming using actual software bugs from open-source projects, Nemotron 3 Ultra achieved a score of 71.7%. This indicates it can generate valid fixes that pass a project's own tests for roughly seven out of ten bugs. NVIDIA stated the benchmark results are verifiable and the post-training methodology is fully disclosed.
NVIDIA also pointed out that the Blackwell platform has already made the high-frequency post-training required for the agentic era economically viable by reducing the cost per run, and the Vera Rubin platform will extend this trajectory further—supporting more rollouts, more parallel environments, and never-ending post-training cycles.
Leading Clients Validate Platform Capabilities
Several companies already running post-training workloads on NVIDIA's platforms have shared specific technical details and expressed intent to migrate to Vera Rubin.
Prime Intellect continuously performs post-training on leading open models using the Blackwell platform and uses NVIDIA Dynamo for inference orchestration. The company has integrated its sandbox infrastructure with the NVIDIA Vera CPU, which showed an average throughput 30% higher than x86 architecture in tests with real RL sandbox workloads. Prime Intellect plans to use Vera Rubin to scale its reinforcement learning environments and accelerate the training-to-inference iteration cycle.
Perplexity's RL post-training stack runs asynchronously across hundreds of NVIDIA GPUs. Its RDMA-based weight transfer engine can synchronize a trillion-parameter model between training and inference nodes within two seconds. The post-trained Qwen3 235B model is subsequently deployed on NVIDIA GB200 NVL72 systems.
Together AI offers post-training capabilities as a service, covering supervised fine-tuning, reinforcement learning, and direct preference optimization, delivered via API and SDK. It currently operates on NVIDIA's platform and has stated it is looking to access the Vera Rubin platform.
Comments